| With the continuous growth of global highway mileage,a large number of roads have entered different maintenance stages,and the amount of maintenance works is also increasing.Therefore,the road technical condition evaluation method needs to be more efficient and accurate.But the traditional artificial pavement condition detection work needs a lot of manpower and material resources,and lacks accuracy and efficiency.Although the image prediction technology based on deep learning has good accuracy,it also has many problems,such as the need for a large number of data labeling and training models,and subjectivity in the process of manually selecting image samples and labeling images.Therefore,this study takes the patched asphalt pavement image as the research object,and through the recognition training of the image,constructs a neural network for predicting the pavement images.At the same time,aiming at the problems of labeling accuracy and training efficiency,an interactive image labeling method based on U-NET convolutional neural network is proposed:(1)Collect and process the pavement image.Image enhancement and denoising based on square equalization and partial differential equation can improve the image quality of asphalt pavement.The image classification algorithm preliminarily screens the repaired images,and completes the construction of neural network training set through semantic segmentation and marking.(2)Construction of asphalt pavement repair image recognition network based on U-Net convolutional neural network model.At the same time,aiming at the problems of heavy marking workload and network learning accuracy,an interactive marking method is designed to optimize the training model of neural network.Interactive marking method is based on the idea of density weight decision of active learning method,and introduces reverse marking to realize the visualization of recognition effect,thus solving the problem of blind network training.The active correction of the reverse labeling results helps U-NET convolutional neural network to actively correct errors,which makes the sample labeling work more selective and purposeful.(3)Evaluate and apply the neural network.According to the Io U,the proportion of boundary grey area and the proportion of identified noise,the neural network is evaluated jointly.The evaluation results show that the accuracy of interactive annotation method based on mean＿Io U index is 6%higher than that of traditional methods under the same sample size and training rounds.The analysis of the noise and boundary accuracy of the prediction results shows that this method eliminates 92% of the noise in the prediction results and improves the boundary accuracy of the prediction results by 14.1%.Therefore,the interactive labeling method proposed in this paper not only improves the learning efficiency of the network,but also greatly improves the accuracy of the prediction results.It can be applied to the automatic detection and calculation of pavement technical conditions,and is a training method that is generally suitable for all kinds of image recognition networks. |
